Applicable Analysis

An International Journal

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/gapa20

Non-isotropic angular Stockwell transform and the

associated uncertainty principles

Firdous A. Shah & Azhar Y. Tantary

To cite this article: Firdous A. Shah & Azhar Y. Tantary (2021) Non-isotropic angular Stockwell
transform and the associated uncertainty principles, Applicable Analysis, 100:4, 835-859, DOI:
10.1080/00036811.2019.1622681

To link to this article: https://doi.org/10.1080/00036811.2019.1622681

Published online: 31 May 2019.

Submit your article to this journal

Article views: 207

View related articles

View Crossmark data

Citing articles: 11 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=gapa20
APPLICABLE ANALYSIS
2021, VOL. 100, NO. 4, 835–859
https://doi.org/10.1080/00036811.2019.1622681

Non-isotropic angular Stockwell transform and the associated

uncertainty principles
Firdous A. Shah and Azhar Y. Tantary
Department of Mathematics, University of Kashmir, Anantnag, India

ABSTRACT ARTICLE HISTORY

For an efficient and directional representation of signals in higher dimen- Received 19 January 2019
sions, we propose the non-isotropic angular Stockwell transform in the Accepted 17 May 2019
context of time-frequency analysis. The proposed transform is aimed at COMMUNICATED BY
rectifying the conventional Stockwell transform by employing an angu- Boris Vainberg
lar and scalable localized window which offers directional flexibility and
thereby results in the multi-scale and directional analysis of signals in KEYWORDS
higher dimensions. The basic properties of the proposed transform such Stockwell transform; window
as orthogonality relation, reconstruction formula, derivation of the admis- function; wavelet transform;
sibility condition and characterization of the range are discussed using discrete Stockwell transform;
Fourier transform;
the machinery of operator theory and Fourier transforms. In addition, we uncertainty principle
introduce the discrete version of the non-isotropic angular Stockwell trans-
form and establish a sufficient condition for the corresponding discrete AMS SUBJECT
family to be a frame in L2 (R2 ). Furthermore, some generalizations of the CLASSIFICATIONS
well-known Heisenberg-type inequalities are derived for the non-isotropic 65R10; 47G10; 42C40; 42C15;
angular Stockwell transform in the Fourier domain. 42A38; 94A12

1. Introduction
Most of the signals in nature are non-stationary and a complete representation of these signals
requires frequency analysis that is local in time, resulting in the time-frequency analysis of signals.
Although time-frequency analysis of signals had its origin almost 60 years ago, there has been a major
development of the time-scale and time-frequency analysis approach in the last few decades and many
new transforms have been introduced to analyze the non-stationary and multi-component signals in
the joint time-frequency domain [1,2]. The conventional short-time Fourier transform (STFT) was
first introduced by Gabor [3] using a Gaussian distribution function as a window function with the
aim of constructing efficient time-frequency localized expansions of finite energy signals f ∈ L2 (R)
as
 ∞
1
Gg [f ](b, ξ ) = √ f (t)ψ(t − b) e−iξ t dt, ξ , b ∈ R. (1)
2π −∞

The idea behind this transform is segmenting a signal f (t) by a rigid time-localized window ψ(t − b)
and then by performing the Fourier transform over these segments. Although this approach has
rectified many limitations of the classical Fourier transform, however, it suffers from the undesir-
able tradeoff between the time concentration and the frequency concentration due to the rigidity of
the window function [4]. To circumvent this limitation, the notion of continuous wavelet transform

CONTACT Firdous A. Shah [email protected]

© 2019 Informa UK Limited, trading as Taylor & Francis Group

836 F. A. SHAH AND A. Y. TANTARY

(CWT) was introduced in the context of time-frequency analysis as

 ∞  
  1 t−b
Wψ f (a, b) = √ f (t)ψ dt, a, b ∈ R, a = 0, (2)
|a| −∞ a

where a is called a scaling parameter which measures the degree of compression or scale, and b is a
translation parameter which determines the time location of the wavelet. Wavelet transforms have
gained a respectable status due to their applications in various disciplines and as such have many
success stories [5–7]. However, it suffers from three major disadvantages: shift sensitivity, poor direc-
tionality and the lack of phase information. Subsequently, some off-shoots of the wavelet transform,
such as the Stockwell transform [8], ridgelet transform [9], curvelet transform [10], contourlet trans-
form [11], shearlet transform [12] and the fractional wavelet transform [13], have been introduced
to address the shortcoming of the wavelet transform. To address the absolute referenced phase infor-
mation of the wavelet transform (2), Stockwell et al. [8] came up with a new mathematical transform
by combining the merits of STFT and wavelet transform called the continuous Stockwell transform
(CST). The modified CST of a finite energy signal f ∈ L2 (R) with respect to a window function
ψ ∈ L2 (R) is defined by
 ∞
  |ξ |  
Sψ f (b, ξ ) = √ e−it·ξ f (t)ψ ξ(t − b) dt, b ∈ R, ξ ∈ R \ {0},
2π −∞

where b and ξ are the time of the spectral localization and the Fourier frequency, respectively. There
are two important terms involved in the modified CST: the phase function e−it·ξ , which provides the
absolutely referenced phase information about √ the phase arg(Sψ f )(b, ξ ) of the given transform at
time b = 0 and the normalization factor |ξ |/ 2π, which normalizes the localizing window to have
unit area. Therefore, CST can be considered as a frequency-dependent STFT or a phase corrected
wavelet transform and thus provides more exact information of local behavior of a signal in the multi-
scale analysis. The original CST introduced by Stockwell et al. [8] can be obtained from (1) by using
Gaussian function e−t /2 as an analyzing window function so that Equation (1) can be recast as
2

 ∞
  |ξ |
e−it·ξ f (t) e−(t−b) ξ /2 dt, b ∈ R, ξ ∈ R \ {0}.
2 2
Sψ f (b, ξ ) = √
2π −∞

One can infer from the above definition that the width of the window function e−(t−b) ξ /2 shrinks
2 2

as the analyzed frequency increases providing a better time localization for high frequencies. It is
pertinent to note that the phase information is provided by referencing the windowed signal f (t)
with respect to e−i(t−b) in case of the so-called Morlet wavelet transform. A detailed comparison
between the CST and the wavelet transform can be found in [14]. In recent years, CST has been
profoundly and intensively investigated because of its important applications in various disciplines
including geophysics, gravitational waves, optics, biomedical imaging, oceanology, bioinformatics,
filtering estimation and signal recovery [15–24].
To harness the advantages of the Stockwell transform for higher dimensional signals, Liu and
Wong [25] extended the classical Stockwell transform to two dimensions. However, due to the direc-
tional insensitivity of the window function, the two-dimensional CST is inadequate for capturing
the anisotropic features like edges and corners in signals. The detection of such geometrical features
in signals is often highly desirable in numerous practical applications including medical imaging,
remote sensing, character recognition and several other areas. The purpose of this paper is to address
this issue by introducing a new time frequency transform namely non-isotropic angular Stockwell
transform (NAST) which employs an angular, scalable and localized window function for capturing
the geometric features in images or two-dimensional signals in general. Unlike the traditional CST,
the proposed transform has the advantage of being directional selective which in turn correspond to
 APPLICABLE ANALYSIS 837

its flexible nature. Moreover, the proposed transform is expected to be a competent and promising
tool for analyzing noisy, multi-component and non-stationary signals wherein directional features
play a role.
The main contributions of this article are as follows:

• to introduce a new time-frequency analysis tool by employing a scalable, angular and localized
window function;
• to derive basic properties of the new transform, such as its orthogonality relation, reconstruction
formula, range theorem and the admissibility condition using the machinery of two-dimensional
Fourier transform;
• to introduce the discrete version of the proposed transform and derive a sufficient condition for
the corresponding discrete family to be a frame for L2 (R2 );
• to derive some generalized Heisenberg-type inequalities for the proposed transform.

The article is organized as follows. In Section 2, we introduce the notion of non-isotropic angular
Stockwell transform and investigate some of its fundamental properties including the orthogonality
relation, reconstruction formula and obtain a complete characterization of the range. In Section 4, we
introduce the discrete version of the non-isotropic angular Stockwell transform and establish a suffi-
cient condition for the corresponding discrete family to be a frame in L2 (R2 ). Section 3 is devoted for
investigating the Heisenberg-type inequalities and some of its generalizations for the non-isotropic
angular Stockwell transform.

2. The non-isotropic angular Stockwell transform

In this section, we shall formally introduce a new transform in the context of time-frequency analysis
namely the non-isotropic angular Stockwell transform. The proposed transform employs an angular,
scalable and localized window function which in turn corresponds to the anisotropic and flexible
nature of the transform.
To establish the mathematical theory for the proposed transform, we shall introduce the elemen-
tary operators on the space L2 (R2 ). For ξ = (ξ1 , ξ2 ) ∈ R2 , b = (b1 , b2 ) ∈ R2 and θ ∈ [0, 2π ), let
DAξ , Eξ , Tb and Rθ be the unitary operators acting on ψ ∈ L2 (R2 ) given by dilation, modulation,
translation and rotation, respectively:

   
DAξ ψ (t) = det Aξ ψ(Aξ t), Eξ ψ (t) = eiξ ·t ψ(t),
(Tb ψ) (t) = ψ(t − b), (Rθ ψ) (t) = ψ(R−θ t),

where
   
ξ 0 cos θ sin θ
Aξ = 1 , ξ1 , ξ2 = 0 and Rθ = .
0 ξ2 − sin θ cos θ

θ (t) in L2 (R2 ) as
With the aid of these operators, we define a family of functions ψb,ξ

θ
 
ψb,ξ (t) := Eξ Rθ DAξ Tb ψ(t) = | det Aξ |ei t·ξ ψ(R−θ Aξ (t − b)), ξ ∈ R2 , b ∈ R2 , θ ∈ 0, 2π . (3)

We are now in a position to introduce the formal definition of the non-isotropic angular Stockwell
transform.
838 F. A. SHAH AND A. Y. TANTARY

Definition 2.1: For any function f ∈ L2 (R2 ), the non-isotropic angular Stockwell transform with
respect to a window function ψ ∈ L2 (R2 ) is denoted by Mψ f and is given by

1 θ (t) dt,
[Mψ f ](b, ξ , θ) = f (t)ψb,ξ (4)
2π R2
θ (t) is given by (3).
where ψb,ξ

It is worth noticing that the proposed transform (4) serves as a local filter analyzing the signal
f (t) over the region that lives around (b, ξ ) in the time-frequency domain about an orientation θ .
For the demonstration of the non-isotropic angular Stockwell transform (4), we shall present some
illustrative examples.

Example 2.2: (i) For the constant function f (t) = M and the non-isotropic Gaussian window func-
θ (t)
tion ψ(t) = eik·t e−(1/2)|t| , k = (k1 , k2 ) ∈ R2 , t = (t1 , t2 ) ∈ R2 , we first construct the family ψb,ξ
2

in the following way:

  
  cos θ − sin θ ξ1 t1 − ξ1 b1
R−θ Aξ (t − b) =
sin θ cos θ ξ2 t2 − ξ2 b2
 
cos θ(ξ1 t1 − ξ1 b1 ) − sin θ(ξ2 t2 − ξ2 b2 )
= .
sin θ(ξ1 t1 − ξ1 b1 ) + cos θ(ξ2 t2 − ξ2 b2 )
Therefore, we have
 

ψ[R−θ Aξ (t − b) ] = exp{ik1 cos θ(ξ1 t1 − ξ1 b1 ) − sin θ(ξ2 t2 − ξ2 b2 )


+ ik2 sin θ(ξ1 t1 − ξ1 b1 ) + cos θ(ξ2 t2 − ξ2 b2 ) }
⎧  2 ⎫
⎪
⎪ cos θ(ξ t − ξ b ) − sin θ(ξ t − ξ b ) ⎪
⎪
⎪
⎪  1 1 1 1 2 2 2 2
 ⎪
⎨ + sin θ(ξ t − ξ b ) + cos θ(ξ t − ξ b ) ⎪ 2 ⎬
1 1 1 1 2 2 2 2
× exp − .
⎪
⎪ 2 ⎪
⎪
⎪
⎪ ⎪
⎪
⎩ ⎭

Hence,
θ
ψb,ξ (t) = | det Aξ | ei t·ξ ψ(R−θ Aξ (t − b))


= | det Aξ | exp{i(t1 ξ1 + t2 ξ2 )} × exp{ik1 cos θ(ξ1 t1 − ξ1 b1 ) − sin θ(ξ2 t2 − ξ2 b2 )


+ ik2 sin θ(ξ1 t1 − ξ1 b1 ) + cos θ(ξ2 t2 − ξ2 b2 ) }
⎧  2 ⎫
⎪
⎪ cos θ(ξ t − ξ b ) − sin θ(ξ t − ξ b ) ⎪
⎪
⎪ 1 1 1 1 2 2 2 2 ⎪
⎨ + sin θ(ξ t − ξ b ) + cos θ(ξ t − ξ b )2 ⎪
⎪ ⎬
1 1 1 1 2 2 2 2
× exp − .
⎪
⎪ 2 ⎪
⎪
⎪
⎪ ⎪
⎪
⎩ ⎭

For simplicity, we shall compute the non-isotropic angular Stockwell transform for the given function
f with respect to the non-isotropic Gaussian function ψ(t) about an orientation of θ = π/2 as
 
M| det Aξ | (ξ1 b1 )2 + (ξ2 b2 )2  
[Mψ f ](b, ξ , π/2) = exp − exp{−i k1 ξ2 b2 − k2 ξ1 b1 }
2π 2
 ∞  
  ξ 2 t2
× exp ξ12 b1 − iξ1 − ik2 ξ1 t1 − 1 1 dt1
−∞ 2
 APPLICABLE ANALYSIS 839

 ∞  
 2
 ξ22 t22
× exp ik1 ξ2 − iξ2 + ξ2 b2 t2 − dt2
−∞ 2
 
M| det Aξ | (ξ1 b1 )2 + (ξ2 b2 )2  
= exp − exp{−i k1 ξ2 b2 − k2 ξ1 b1 }
ξ1 ξ2 2
 2  2 
ξ12 b1 − iξ1 − ik2 ξ1 ik1 ξ2 − iξ2 + ξ22 b2
× exp + .
2ξ12 2ξ22

In particular, for b = (0, 0) and ξ = (1, 1), we have

 
2k1 − 2k2 − k21 − k22 − 2
[Mψ f ](0, 1, π/2) = M exp .
2

Similarly, for the orientation θ = π, we obtain

 
M| det Aξ | (ξ1 b1 )2 + (ξ2 b2 )2  
[Mψ f ](b, ξ , π/2) = exp − exp{i k1 ξ1 b1 + k2 ξ2 b2 }
ξ1 ξ2 2
 2  2 2 
ξ1 b1 − iξ1 + ik1 ξ1
2 ξ2 b2 + ik2 ξ2 − iξ2
× exp + .
2ξ12 2ξ22

(ii) We now consider the isotropic Gaussian window function ψ(t) = Ce−|t|
2 /2
, C ∈ R, t =
(t1 , t2 ) ∈ R2 , then we have the following collection of analyzing functions:
θ
ψb,ξ (t) = | det Aξ | ei t·ξ ψ(R−θ Aξ (t − b))
⎧  2 ⎫
⎪
⎪ cos θ(ξ t − ξ b ) − sin θ(ξ t − ξ b ) ⎪
⎪
⎪ 1 1 1 1 2 2 2 2 ⎪
⎨ + sin θ(ξ t − ξ b ) + cos θ(ξ t − ξ b )2 ⎪
⎪ ⎬
1 1 1 1 2 2 2 2
= C| det Aξ | exp{i(t1 ξ1 + t2 ξ2 )} exp − .
⎪
⎪ 2 ⎪
⎪
⎪
⎪ ⎪
⎪
⎩ ⎭

Here we compute the non-isotropic angular Stockwell transform for the constant function f (t) = M
with respect to the Gaussian function ψ(t) at θ = π/2 as
 
CM| det Aξ | (ξ1 b1 )2 + (ξ2 b2 )2
[Mψ f ](b, ξ , π/2) = exp −
2π 2
 ∞
× exp{(ξ12 b1 − iξ1 )t1 − (ξ12 /2)t12 } dt1
−∞
 ∞
× exp{(ξ22 b2 − iξ2 )t2 − (ξ22 /2)t22 } dt2
−∞
 
CM| det Aξ | (ξ1 b1 )2 + (ξ2 b2 )2
= exp −
ξ1 ξ2 2
 2  2 
ξ12 b1 − iξ1 ξ22 b2 − iξ2
× exp + .
2ξ12 2ξ22

In particular, for b = (0, 0) and ξ = (1, 1), we have

[Mψ f ](0, 1, π/2) = CM/e.

840 F. A. SHAH AND A. Y. TANTARY

(iii) Consider the modulation function f (t) = eit·ω , t, ω ∈ R2 and let ψ(t) be as in preceding case (ii),
then we can compute the non-isotropic angular Stockwell transform of f (t) at θ = π/2 as
 
| det Aξ | (ξ1 b1 )2 + (ξ2 b2 )2
[Mψ f ](b, ξ , π/2) = exp −
2π 2
 ∞
× exp{(ξ12 b1 − i(ξ1 − ω1 ))t1 − (ξ12 /2)t12 } dt1
−∞
∞
× exp{(ξ22 b2 − i(ξ2 − ω2 ))t2 − (ξ22 /2)t22 } dt2
−∞
 
| det Aξ | (ξ1 b1 )2 + (ξ2 b2 )2
= exp −
ξ1 ξ2 2
 2  2 2 
ξ12 b1 − i(ξ1 − ω1 ) ξ2 b2 − i(ξ2 − ω2 )
× exp + .
2ξ12 2ξ22

In particular, for b = (0, 0) and ξ = (1, 1), we have

 
(1 − ω1 )2 + (1 − ω2 )2
[Mψ f ](0, 1, π/2) = exp − .
2

Following theorem assembles some basic properties of the proposed transform (4).

Theorem 2.3: For any f , g ∈ L2 (R2 ), α, β ∈ C and k ∈ R2 , λ ∈ R+ , the non-isotropic angular Stock-
well transform (4) satisfies the following properties:

(i) Linearity: [Mψ (αf + βg)](b, ξ , θ) = α[Mψ f ](b, ξ , θ) + β[Mψ g](b, ξ , θ),
(ii) Anti-linearity: [Mαψ+βφ f ](b, ξ , θ) = α [Mψ f ](b, ξ , θ) + β [Mφ f ](b, ξ , θ),
(iii) Translation: [Mψ (Tk f )](b, ξ , θ) = e−ik·ξ [Mψ f ](b − k, ξ , θ),
(iv) Scaling: [Mψ (f (λ t))](b, ξ , θ) = [Mψ f ](λ b, λξ , θ),
(v) Reflection: [Mψ (f (−t))](b, ξ , θ) = −[Mψ f ](−b, −ξ , θ),
(vi) Conjugation: [Mψ f ](b, ξ , θ) = [M ˇ f ](b, −ξ , θ), ψ̌(t) = ψ(−t).
ψ̄

Proof: (i). For α, β ∈ C, we have

 
θ
[Mψ (αf + βg)](b, ξ , θ) = αf + βg, ψb,ξ
   
θ θ
= αf , ψb,ξ + βg, ψb,ξ
   
= α Mψ f (b, ξ , θ) + β Mψ g (b, ξ , θ).

(ii). Let ψ, φ ∈ L2 (R2 ) be given window functions, then for the scalars α, β ∈ C, we have
   
Mαψ+βφ f (b, ξ , θ) = f , {αψ + βφ}θb,ξ
   
θ θ
= α f , ψb,ξ + β g, φb,ξ
   
= α Mψ f (b, ξ , θ) + β Mφ f (b, ξ , θ).
 APPLICABLE ANALYSIS 841

(iii). For any k ∈ R2 , we have

 
| det Aξ |  
[Mψ (Tk f )](b, ξ , θ) = e−it·ξ f (t − k)ψ R−θ Aξ (t − b) dt
2π R2
 
| det Aξ |  
= e−i(z+k)·ξ f (z)ψ R−θ Aξ (z + k − b) dz
2π R2
 
| det Aξ | e−ik·ξ  
= e−i z·ξ f (z)ψ R−θ Aξ (z − (b − k)) dz
2π R2
 
= e−ik·ξ Mψ f (b − k, ξ , θ).

(iv). For every λ ∈ R+ , we observe that

 
  | det Aξ |   
Mψ f (λ t) (b, ξ , θ) = e−i t·ξ f (λ t)ψ R−θ Aξ (t − b) dt
2π R2
   z  
| det Aξ |
= e−i z·ξ/λ f (z)ψ R−θ Aξ −b dz
2π λ R2 λ
  
| det Aξ |  
= e−iz·ξ/λ f (z) ψ R−θ Aξ/λ (z − λb) dz
2π λ R2
  
ξ
= Mψ f λb, , θ .
λ

(v). Finally, we observe that

 
| det Aξ |  
[Mψ f (−t)](b, ξ , θ) = e−it·ξ f (−t)ψ R−θ Aξ (t − b) dt
2π R2

| det Aξ |   
=− e−i z·(−ξ ) f (z)ψ R−θ Aξ (−z − b) dz
2π R2
 
| det Aξ |  
=− e−i z·(−ξ ) f (z)ψ R−θ A−ξ (z + b) dz
2π R2
 
= − Mψ f (−b, −ξ , θ).

(vi). We have

| det Aξ |   
[Mψ f ](b, ξ , θ) = e−i t·ξ f (t)ψ R−θ Aξ (t − b) dt
2π R2
 
| det Aξ |  
= e−i t·(−ξ ) f (t)ψ R−θ Aξ (t − b) dt
2π R2
 
| det A−ξ |  
= e−i t·(−ξ ) f (t)ψ − R−θ A−ξ (t − b) dt
2π R2
 
= M ˇ f (b, −ξ , θ).
ψ̄

This completes the proof of Theorem 2.3. 

We now establish a pivotal lemma which expresses a relationship between the proposed transform
(Mψ f )(b, ξ , θ) and the Fourier transform of the individual functions f (t) and ψ(t).
842 F. A. SHAH AND A. Y. TANTARY

Lemma 2.4: Let [Mψ f ](b, ξ , θ) be the non-isotropic angular Stockwell transform of the function f ∈
L2 (R2 ), then we have
    
−iξ ·b −1 −1
Mψ f (b, ξ , θ) = e F f̂ (ω)ψ̂ (ω − ξ )Aξ Rθ (b), (5)

where F denotes the two-dimensional Fourier transform of f (t) given by


1
F [f ](ω) = f̂ (ω) = f (t) e−i t·ω dt. (6)
2π R2

Proof: For any f (t) ∈ L2 (R2 ), we have


θ 1
F [ψb,ξ ](ω) = ψ θ (t) e−i t·ω dt
2π R2 b,ξ

| det Aξ |  
= ei t·ξ ψ[R−θ Aξ (t − b) ] e−it·ω dt
2π R 2
  
| det Aξ |
= ei(z+b)·ξ ψ R−θ Aξ z e−i(z+b)·ω dz
2π R2

1
= ei(ξ −ω)·b ψ(y) exp{i(ξ − ω) · (A−1
ξ Rθ y)} dy
2π R2

ei(ξ −ω)·b
= ψ(y) exp{−i((ω − ξ )A−1 ξ Rθ ) · y} dy
2π R2
  
= ei(ξ −ω)·b F ψ (ω − ξ )A−1 ξ Rθ . (7)

By virtue of the expression (7), we compute the proposed transform (Mψ f )(b, ξ , θ) as

1  θ

[Mψ f ](b, ξ , θ) = f , ψb,ξ
2π
1  
θ

= F [f ], F ψb,ξ
2π
   
1
= F [f ] e−i(ξ −ω)·b F ψ (ω − ξ )A−1 ξ Rθ dω
(2π ) R2
   
e−iξ ·b
= F [f ]F ψ (ω − ξ )A−1 ξ R θ e
iω·b
dω
(2π ) R2
  
−iξ ·b −1 −1
=e F f̂ (ω)ψ̂ (ω − ξ )Aξ Rθ (b).

This completes the proof of Lemma 2.4. 

Observation. The following identity is obtained by taking modulus on both sides of (7)

   
F ψ θ (ω) = F ψ (ω − ξ )A−1 Rθ . (8)
b,ξ ξ

Equation (8) shall play an important role for deriving the upcoming admissibility condition (9)
associated with the pre-defined transform (4).
 APPLICABLE ANALYSIS 843

Theorem 2.5: Any window function ψ ∈ L2 (R2 ) is admissible if

     2
2π
Cψ = F ψ (ω − ξ )A−1 Rθ dθ dξ < ∞, a.e. (9)
ξ
R2 \{0} 0

Proof: We know that the window function ψ ∈ L2 (R2 ) is admissible if [2]

   2π   2
θ
ψ, ψb,ξ dθ dξ db < ∞. (10)
R2 R2 \{0} 0

By virtue of Lemma 2.4 and the Schwartz’s inequality, we obtain

   2π   2    2π   2
θ θ
ψ, ψb,ξ dθ dξ db = F [ψ], F [ψb,ξ ] dθ dξ dω
R2 R2 \{0} 0 R2 R2 \{0} 0
   2π 2   2


≤ F [ψ](ω) F ψb,ξ (ω) dθ dξ dω
θ
R2 R2 \{0} 0
   2π 2

= F [ψ](ω)
R2 R2 \{0} 0
   2
× F ψ (ω − ξ )A−1 R
θ dθ dξ dω
ξ
 2   2π

= F [ψ](ω)
R2 R2 \{0} 0
  
 2
× F ψ (ω − ξ )Aξ Rθ dθ dξ dω,
−1

which is finite, provided

     2
2π
Cψ = F ψ (ω − ξ )A−1 Rθ dθ dξ < ∞.
ξ
R2 \{0} 0

This completes the proof of Theorem 2.5. 

In our next theorem, we shall establish an orthogonality formula between two signals and their
respective non-isotropic angular Stockwell transforms. As a consequence of this formula, we can
deduce the resolution of identity for the proposed transform (4).

Theorem 2.6: If ψ ∈ L2 (R2 ) and [Mψ f ](b, ξ , θ) is the non-isotropic angular Stockwell transform of f
defined by (4), then for any f , g ∈ L2 (R2 ), we have

      
2π 
Mψ f (b, ξ , θ) Mψ g (b, ξ , θ) dθ dξ db = Cψ f , g . (11)
R2 R2 \{0} 0
844 F. A. SHAH AND A. Y. TANTARY

Proof: Applying the Parseval’s formula for the two-dimensional Fourier transform, the non-isotropic
angular Stockwell transform (4) of any function f ∈ L2 (R2 ) can be expressed as

1 θ
[Mψ f ](b, ξ , θ) = f , ψb,ξ 
2π
! "
1 θ
= F [f ], F [ψb,ξ ]
2π
   
1
= F [f ](ω) e−i(ξ −ω)·b F ψ (ω − ξ )A−1
ξ Rθ dω.
(2π ) R2

A similar computation for g ∈ L2 (R2 ) yields

     
1
Mψ g (b, ξ , θ) = F [g](η) e−i(ξ −η)·b F ψ (η − ξ )A−1
ξ R θ dη.
(2π ) R2

Therefore, by virtue of Fubini theorem, we can write

   2π    
Mψ f (b, ξ , θ) Mψ g (b, ξ , θ) dθ dξ db
R2 R2 \{0} 0
   2π     
1 −i(ξ −ω)·b −1
= F [f ](ω) e F ψ (ω − ξ )Aξ R θ dω
(2π )2 R2 R2 \{0} 0 R2
    
i(ξ −η)·b −1
× F [g](η) e F ψ (η − ξ )Aξ Rθ dη dθ dξ db
R2
    2π   
= F [f ](ω)F [g](η)F ψ (η − ξ )A−1
ξ Rθ
R2 R2 R2 \{0} 0
   1
 
× F ψ (η − ξ )A−1
ξ R θ ei(ω−η)·b db dθ dξ dω dη
(2π )2 R2
      2
2π
= F [f ](ω)F [g](ω) F ψ (ω − ξ )A−1
ξ R
θ dθ dξ dω
R2 R2 \{0} 0
   
  2π  2
= F [f ](ω)F [g](ω) F ψ (ω − ξ )A−1 Rθ dθ dξ dω
ξ
R2 R2 \{0} 0

= Cψ F [f ](ω)F [g](ω) dω
R2
 
= Cψ F [f ], F [g]

= Cψ f , g .

This completes the proof of Theorem 2.6. 

Corollary 2.7: For f = g, the orthogonality relation (11) yields

#  #2 # #2
# #
# Mψ f (b, ξ , θ)# = Cψ # #
#f # . (12)
# # 2
 APPLICABLE ANALYSIS 845

Remarks: (i) For any window function ψ ∈ L2 (R2 ) such that Cψ = 1, the non-isotropic angular
Stockwell transform becomes an isometry from the space of signals to the space of transforms
or more precisely from L2 (R2 ) to L2 (R2 × R2 \ {0} × [0, 2π )).
(ii) The operator Mψ f is one–one from L2 (R2 ) to L2 (R2 × R2 \ {0} × [0, 2π )).

The next theorem guarantees the reconstruction of the input signal from the corresponding non-
isotropic angular Stockwell transform (4).

Theorem 2.8: Any signal f ∈ L2 (R2 ) can be reconstructed from the non-isotropic angular Stockwell
transform [Mψ f ](b, ξ , θ) via the following formula:

   2π  
1 θ
f (t) = Mψ f (b, ξ , θ) ψb,ξ (t) dθ dξ db a.e. (13)
2πCψ R2 R2 \{0} 0

Proof: Invoking the orthogonality formula (11) for any arbitrary function g ∈ L2 (R2 ), we obtain

      
 2π
Cψ f , g = Mψ f (b, ξ , θ) Mψ g (b, ξ , θ) dθ dξ db
R2 R2 \{0} 0
    
1 2π  
= Mψ f (b, ξ , θ) θ (t) dt dθ dξ db
g(t)ψb,ξ
2π R2 R2 \{0} 0 R2
    2π  
1 θ
= Mψ f (b, ξ , θ)ψb,ξ (t) g(t) dθ dξ db dt
2π R2 R2 R2 \{0} 0
!   2π   "
1 θ
= Mψ f (b, ξ , θ) ψb,ξ (t) dθ dξ db, g(t) .
2π R2 R2 \{0} 0

Since g is arbitrarily, therefore it follows that

   2π  
1 θ
f (t) = Mψ f (b, ξ , θ) ψb,ξ (t) dθ dξ db a.e.
2πCψ R2 R2 \{0} 0

This completes the proof of Theorem 2.8. 

In the upcoming theorem, we present a complete characterization of the range of the transform
[Mψ f ](b, ξ , θ). The result follows as a consequence of the reconstruction formula (13) and the well-
known Fubini theorem.

Theorem 2.9: Let ψ ∈ L2 (R2 ) be a window function with Cψ = 1. Then, a function f ∈ L2 (R2 ×
R2 \ {0} × [0, 2π]) is the non-isotropic angular Stockwell transform of a certain signal if and only if it
satisfies the reproduction property:

    

 1 2π
θ θ
f c, η, θ = f (b, ξ , θ) ψb,ξ , ψc,η dθ dξ db. (14)
(2π )2 R2 R2 \{0} 0
846 F. A. SHAH AND A. Y. TANTARY

Proof: Let f ∈ Mψ (L2 (R2 )), then there exists a function g ∈ L2 (R2 ), such that Mψ g = f . We show
that f satisfies (14). To show this, we proceed as
   
f c, η, θ = Mψ g (c, η, θ )

1 θ (t) dt
= g(t) ψc,η
2π R2
     2π   
1 1 θ θ (t) dt
= Mψ g (b, ξ , θ)ψb,ξ (t) dθ dξ db ψc,η
2π R2 2πCψ R2 R2 \{0} 0
   2π    
1 θ θ (t) dt dθ dξ db
= M ψ g (b, ξ , θ) ψ b,ξ (t) ψc,η
(2π )2 R2 R2 \{0} 0 R2
   2π  
1 θ θ
= f (b, ξ , θ) ψ b,ξ , ψc,η dθ dξ db.
(2π )2 R2 R2 \{0} 0

Conversely, suppose that a function f ∈ L2 (R2 × R2 \ {0} × [0, 2π]) satisfies (14). We show that f ∈
Mψ (L2 (R2 )). That is, we shall obtain a function g ∈ L2 (R2 ) such that Mψ g = f . We claim that
   2π
1 θ
g(t) = f (b, ξ , θ) ψb,ξ (t) dθ dξ db.
2π R2 R2 \{0} 0

First, we show that g ∈ L2 (R2 ). Therefore, we have


# #2
#g # = g(t)g(t) dt
2
R2
     2π 
1 θ
= f (b, ξ , θ) ψb,ξ (t) dθ dξ db
R2 2π R2 R2 \{0} 0
    2π 
1 
 θ
× f c, η, θ ψc,η (t) dθ dη dc
2π R2 R2 \{0} 0
   2π    2π   
1 θ θ
= f (b, ξ , θ) ψb,ξ , ψc,η dθ dξ db
(2π )2 R2 R2 \{0} 0 R2 R2 \{0} 0
 
× f c, η, θ dθ dη dc
   2π
   
= f c, η, θ f c, η, θ dθ dη dc
R2 R2 \{0} 0
# #2
= #f #2 .
Consequently, we have

  1
Mψ g (c, η, θ ) = θ (t) dt
g(t) ψc,η
2π R2
    2π 
1 θ
θ (t) dt
= f (b, ξ , θ) ψb,ξ (t) dθ dξ db ψc,η
(2π )2 R2 R2 R2 \{0} 0
   2π  
1 θ θ
= f (b, ξ , θ) ψb,ξ , ψc,η dθ dξ db
(2π )2 R2 R2 \{0} 0

= f (c, η, θ ).
This completes the proof of Theorem 2.9. 
 APPLICABLE ANALYSIS 847

Corollary 2.10: Let ψ ∈ L2 (R2 ) be a window function with Cψ = 1, then Mψ (L2 (R2 )) is a repro-
ducing kernel Hilbert space in L2 (R2 × R2 \ {0} × [0, 2π )) with the kernel given by
  1  θ θ

Kψα b, ξ , θ; c, η, θ = ψ b,ξ , ψc,η . (15)
(2π )2

3. Heisenberg-type inequalities for the non-isotropic angular Stockwell transform

Time-frequency analysis plays a pivotal role in situations where the time dependence of the analyzed
signal is of the same importance as its frequency dependence. However, the optimal time-frequency
resolution of a signal is restricted to a certain limit as stated by the classical Heisenberg’s uncertainty
principle in classical harmonic analysis. Precisely, it states that ‘a function cannot be sharply localized
in both the time and frequency domains’. That is, increasing the knowledge of the time decreases the
knowledge of the frequency of signal and vice versa. The essence of the uncertainty principle is that
it provides a lower bound for optimal resolution of a signal in both the time and frequency domains.
The classical Heisenberg–Pauli–Weyl uncertainty inequality for any f ∈ L2 (R2 ) is
 1/2  2 1/2 1  
2
|t|2 f (t) dt |ξ |2 f̂ (ξ ) dξ ≥ f (t) 2 dt
R2 R2 2 R2

with equality if and only if f is a multiple of a suitable Gaussian function. This standard inequality
has been extended in different settings and for various transforms including windowed Fourier trans-
form, wavelet transform, Wigner–Ville transform, fractional Fourier transform, fractional wavelet
transform, Hankel transform, Dunkl transform, shearlet transform and many more [26–39].
In this section, we shall establish some new versions of the Heisenberg’s uncertainty principle
associated with the proposed non-isotropic angular Stockwell transform. First, we shall derive the
Heisenberg-type inequality by regarding the proposed transform as a function of time variable b
and then its counterpart in the frequency variable ξ . Subsequently, we shall obtain the phase space
uncertainty principle for the proposed transform in the form of Theorem 3.3. Finally, we shall derive
a logarithmic version of the Heisenberg-type inequality by virtue of Pitt’s inequality.

Theorem 3.1: If [Mψ f ](b, ξ , θ) is the non-isotropic angular Stockwell transform of any nontrivial
function f ∈ L2 (R2 ) with respect to the window function ψ, then the following uncertainty inequality
holds:
   2π   2 1/2  2 1/2 √C # #2
2 2 ψ # #
|b| Mψ f (b, ξ , θ) dθ dξ db |ω| F [f ](ω) dω ≥ #f # .
R2 R2 \{0} 0 R2 2
(16)

Proof: The classical Heisenberg–Pauli–Weyl inequality in the Fourier domain is given by

 1/2  1/2   
2 2 2 2
b f (b) 2 db ω f̂ (ω) dω ≥
1
f (b) db . (17)
R2 R2 2 R2

We identify [Mψ f ](b, ξ , θ) as a function of the translation parameter b and replace f (b) by
[Mψ f ](b, ξ , θ) in (17) to obtain
 2 1/2  2 1/2
2   2   
b Mψ f (b, ξ , θ) db ω F Mψ f (b, ξ , θ) (ω) dω
R2 R2
  2
1 
≥ Mψ f (b, ξ , θ) db. (18)
2 R2
848 F. A. SHAH AND A. Y. TANTARY

Integrating (18) with respect to the measure dξ dθ and using the orthogonality relation (11), we have
   2 1/2   2 1/2
2π 2   2  
b Mψ f (b, ξ , θ) db ω F Mψ f (b, ξ , θ) (ξ ) dω dθ dξ
R2 \{0} 0 R2 R2
   2π  2
1  Cψ # #
#f #2 .
≥ Mψ f (b, ξ , θ) dθ dξ db = (19)
2 R2 R2 \{0} 0 2

By virtue of the Schwartz’s inequality, we can rewrite (19) as

   1/2
2π 2   2
b Mψ f (b, ξ , θ) dθ dξ db
R2 R2 \{0} 0
   1/2
2π 2    2 Cψ # #
× ω F Mψ f (b, ξ , θ) (ω) dθ dξ dω ≥ #f #2 . (20)
R2 R2 \{0} 0 2

Using Lemma 2.4 and the translation invariance of the Lebesgue integral, we observe that
   2π 2    2
ω F Mψ f (b, ξ , θ) (ω) dθ dξ dω
R2 R2 \{0} 0
      2
2  
2π
= ω F f (ω) F ψ (ω − ξ )A−1 Rθ dθ dξ dω
ξ
R2 R2 \{0} 0
  2π    2 
2   2
= F ψ (ω − ξ )A−1 Rθ dθ dξ ω F f (ω) dω
ξ
R2 \{0} 0 R2

2   2
= Cψ ω F f (ω) dω. (21)
R2

Using the above estimate in inequality (20), we get the desired result as follows:
   2π 2   2 1/2 
2 2 1/2 √C # #2
b Mψ f (b, ξ , θ) dθ dξ db ω F [f ](ω) dω ≥
ψ # #
#f # .
R2 R2 \{0} 0 R2 2

This completes the proof of Theorem 3.1. 

The following result is an analogous version of Theorem 3.1 in the frequency variable ξ . The proof
can be obtained in a similar fashion as in preceding theorem.

Theorem 3.2: If [Mψ f ](b, ξ , θ) is the non-isotropic angular Stockwell transform of any nontrivial
function f ∈ L2 (R2 ) with respect to the window function ψ, then the following uncertainty inequality
holds:
   2π
2   2 1/2 
2 2 1/2 √C # #2
ξ Mψ f (b, ξ , θ) dθ dξ db t f (t) dt ≥
ψ # #
#f # . (22)
R2 R2 \{0} 0 R2 2 2

As a consequence of Theorems 3.1 and 3.2, we shall obtain the following inequality which shows
that the phase-space localization of the pair (f , f̂ ) varies inversely with respect to the phase-space
localization of the proposed transform (4), that is the better the phase-space localization of the pair
(f , f̂ ), the worse is the phase-space localization of [Mψ f ](b, ξ , θ) and vice versa.
 APPLICABLE ANALYSIS 849

Theorem 3.3: If [Mψ f ](b, ξ , θ) is the non-isotropic angular Stockwell transform of any non-trivial
function f ∈ L2 (R2 ), then the following uncertainty inequality holds:

   1/2
2π 2   2
b Mψ f (b, ξ , θ) dθ dξ db
R2 R2 \{0} 0
   1/2
2π 2   2
× ξ Mψ f (b, ξ , θ) dθ dξ db
R2 R2 \{0} 0
 2 1/2  2 1/2 C # #4
2 2
× ω F [f ](ω) dω t f (t) dt ≥
ψ # #
#f # . (23)
R2 R2 4 2

Following the idea of Cowling and Price [40], we shall derive a generalization of Theorem 3.1 for
the space Lp (R2 ), 1 ≤ p ≤ 2 and p ≥ 2 in the following theorem.

Theorem 3.4: Let ψ ∈ L2 (R2 ) be a window function which satisfies the admissibility condition defined
by (9). Then, for arbitrary f ∈ L2 (R2 ) \ {0}, we have

   1/p  p 1/p
2π
  p

(i) bMψ f b, ξ , θ dθ db dξ ω F [f ](ω) dω
R2 \{0} R2 0 R2
√
Cψ # #
#f #2 , for 1 ≤ p ≤ 2.
≥ 2
(24)
2
   2π 1/p  2 1/p
p   2 p  
(ii) b Mψ f b, ξ , θ dθ db dξ ω F f (ω) dω
R2 \{0} R2 0 R2
1/p
Cψ # #4/p
# #
≥ #f # , for p ≥ 2. (25)
2 2

Proof: (i) Applying Theorem 1.2 of Cowling and Price [40] for the function Mψ f (b, ξ , θ) in the time
variable b, we obtain
 1/p  1/p 
  p    p 1   2
bMψ f b, ξ , θ db |ωF Mψ f b, ξ , θ )(ω)| dω ≥ Mψ f b, ξ , θ db.
R2 R2 2 R2

Integrating the above inequality with respect to the measure dθ dξ and using the orthogonality
relation (11), we obtain
   1/p  1/p
2π
  p    p
bMψ f b, ξ , θ db |ωF Mψ f b, ξ , θ )(ω)| dω dθ dξ
R2 \{0} 0 R2 R2
  
1 2π
  2 Cψ # #
# #2
≥ Mψ f b, ξ , θ dθ db dξ = #f # . (26)
2 R2 \{0} R2 0 2 2

Using Lemma 2.4 and the translation invariance of Lebesgue integral, we can write
    p
  
|ωF Mψ f b, ξ , θ (ω)|p dω = ωf̂ (ω) ψ̂ (ω − ξ )A−1 Rθ dω,
ξ
R2 R 2
850 F. A. SHAH AND A. Y. TANTARY

which on integration together with the admissibility condition (9) yields

   2π  p
 
|ωF Mψ f b, ξ , θ )(ω) dθ dω dξ
R2 \{0} R2 0
   2π   p

= ψ̂ (ω − ξ )A−1 p
ξ Rθ |ωf̂ (ω)| dθ dω dξ
R2 \{0} R2 0
  2π   2 p/2  

≤ ψ̂ (ω − ξ )A−1
ξ R θ dθ dξ |ωf̂ (ω)|p dω
R2 \{0} 0 R2

p/2
= Cψ |ωf̂ (ω)|p dω. (27)
R2

Using inequalities (26), (27) and the Schwartz’s inequality, we obtain

   1/p  p 1/p
2π
  p

bMψ f b, ξ , θ dθ db dξ ω F [f ](ω) dω
R2 \{0} R2 0 R2
   1/p
1 2π
  p
≥ 1/2 bMψ f b, ξ , θ dθ db dξ
Cψ R2 \{0} R2 0
    p 1/p
2π  
× |ωF Mψ f b, ξ , θ )(ω) dθ db dξ
R2 \{0} R2 0
   1/p
1 2π
  p
≥ 1/2 bMψ f b, ξ , θ db
Cψ R2 \{0} 0 R2
  p 1/p
 
× |ωF Mψ f b, ξ , θ )(ω) dω dθ dξ
R2
√
Cψ # #
#f #2 ,
≥ 2
1 ≤ p ≤ 2.
2
(ii) We now derive another generalized inequality for the proposed transform (4) for the case p ≥ 2.
By virtue of Hölder’s inequality, we can write
   2/p
2π
p
  2
|b| Mψ f b, ξ , θ dθ db dξ
R2 \{0} R2 0
   1−2/p
2π
  2
× Mψ f b, ξ , θ dθ db dξ
R2 \{0} R2 0
   2/p
2π
2
  4/p p/2
= (|b| Mψ f b, ξ , θ ) dθ db dξ
R2 \{0} R2 0
   1−2/p
2π
  2−4/p 1/(1−2/p)
× ( Mψ f b, ξ , θ ) dθ db dξ
R2 \{0} R2 0
   
2π  4/p   2−4/p
≥ |(|b|2 Mψ f b, ξ , θ )( Mψ f b, ξ , θ )| dθ db dξ
R2 \{0} R2 0
  
2π
  2
= bMψ f b, ξ , θ dθ db dξ .
R2 \{0} R2 0
 APPLICABLE ANALYSIS 851

Therefore, we have
   1/p
2π
  2
|b|p Mψ f b, ξ , θ dθ db dξ
R2 \{0} R2 0
$ $ $ 2π  
{ R2 \{0} R2 0 |b Mψ f b, ξ , θ |2 dθ db dξ }1/2
≥ $ $ $ 2π . (28)
{ R2 \{0} R2 0 |Mψ f (b, ξ , θ)|2 dθ db dξ }1/2−1/p

In analogy with above and by virtue of orthogonality relation (11), we have

%$ &1/2 %$ &1/2
 1/p |ωf̂ (ω)|2 dω |ω f̂ (ω)| 2 dω
R2 1/2−1/p R 2
|ω|p |f̂ (ω)|2 dω ≥ %$ &1/2−1/p = Cψ % $ &1/2−1/p
R2
R2 |f̂ (ω)| dω Cψ R2 |f̂ (ω)|2 dω
2

%$ &1/2
|ωf̂ (ω)|2 dω
R2
1/2−1/p
= Cψ  1/2−1/p . (29)
$ $ $ 2π   2
R2 \{0} R2 0 Mψ f b, ξ , θ dθ db dξ

Multiplying the inequalities (28) and (29) and employing Theorem 3.1, we obtain
   1/p  1/p
2π p   2
b Mψ f b, ξ , θ dθ dξ db |ω|p |f̂ (ω)|2 dω
R2 \{0} R2 0 R2
 1/2 %
$ $ $ 2π   2 $ &1/2
R2 \{0} R2 0 bM ψ f b, ξ , θ dθ dξ db R2 |ωf̂ (ω)| dω
2
1/2−1/p
≥ Cψ  1−2/p
$ $ $ 2π   2
R2 \{0} R2 0 Mψ f b, ξ , θ dθ dξ db
# #2 √
#
1/2−1/p #
Cψ # #
1/p
Cψ #f # Cψ # #4/p
= # #2 = #f # .
# # 2 2
2(Cψ #f # )1−2/p

The proof of Theorem 3.4 is complete. 

It is worth noticing that for p = 2, Theorem 3.4 boils down to Theorem 3.1.
In our next theorem, we establish the logarithmic uncertainty inequality associated with the pro-
posed transform (4). Such an inequality is obtained by using the classical Pitt’s inequality [26] and
Lemma 2.4.

Theorem 3.5: For any f ∈ L2 (R2 ), the non-isotropic angular Stockwell transform satisfies the following
logarithmic estimate of the uncertainty inequality:
   2π 2    2

ln |b| Mψ f (b, ξ , θ) dθ dξ db + Cψ ln |ω| F f (ω) dω
R2 R2 \{0} 0 R2
# #2   (1/2) 
# #
≥ Cψ #f # − ln π , (30)
(1/2)

provided the integral on the left-hand side is defined, where (t) denotes the Euler’s gamma function.
852 F. A. SHAH AND A. Y. TANTARY

Proof: The logarithmic uncertainty principle in the classical Fourier domains is given by [26]
  2  
2  (1/2)
ln |b| f (b) db + ln |ω| F [f ](ω) dω ≥ − ln π f (b) 2 db. (31)
R2 R2 (1/2) R2

By the same argument as in Theorem 3.1, we can replace the function f in (31) by Mψ f (b, ξ , θ) so
that after integration with respect to the measure dθ dξ the above inequality takes the form
   2π 2

ln |b| Mψ f (b, ξ , θ) dθ dξ db
R2 R2 \{0} 0
   
2π
 2
+ ln |ω| F Mψ f (b, ξ , θ) (ω) dθ dξ dω
R2 R2 \{0} 0
    2
 (1/2) 2π

≥ − ln π Mψ f (b, ξ , θ) da db. (32)
(1/2) R2 R2 \{0} 0

Using Lemma 2.4 together with the translation invariant property of the Lebesgue integral, the second
integral on the left-hand side (32) simplifies to
   2π   2

ln |ω| F Mψ f (b, ξ , θ) (ω) dθ dξ dω
R2 R2 \{0} 0
   2π    
= ln |ω|F Mψ f (b, ξ , θ) (ω)F Mψ f (b, ξ , θ) (ω) dθ dξ dω
R2 R2 \{0} 0
       
2π  
= ln |ω|F f (ω)F ψ (ω − ξ )A−1
ξ Rθ F f (ω)
R2 R2 \{0} 0
  
× F ψ (ω − ξ )A−1 ξ Rθ dθ dξ dω
  2π   
   2   2
−1
= ln |ω| F f (ω) F ψ (ω − ξ )Aξ Rθ dθ dξ dω
R2 R2 \{0} 0
   2

= Cψ ln |ω| F f (ω) dω.
R2

Using the above estimate in (32) and noting that ψ is an admissible window function, we obtain
   2π    2
2

ln |b| Mψ f (b, ξ , θ) dθ dξ db + Cψ ln |ω| F f (ω) dω
R2 R2 \{0} 0 R2
# #2   (1/2) 
# #
≥ Cψ #f # − ln π .
2 (1/2)

This completes the proof of Theorem 3.5. 

4. Discrete non-isotropic angular Stockwell transform

In this section, we introduce a discrete analogue of the proposed non-isotropic angular Stock-
well transform (4). The procedure relies on an appropriate discretization of the parameters b, ξ
θ (t) and the composition of the involved parameters is to be car-
and θ involved in the family ψb,ξ
ried out in a well-defined manner. For every m0 , m1 ∈ Z
 and λ > 1, we discretize the frequency
parameter ξ = (ξ1 , ξ2 ) as ξ j = Bj (m0 , m1 ), where Bj = λ−j 0 , so that ξj = Bj (1, 1) for m0 =
0 λ−j
 APPLICABLE ANALYSIS 853

m1 = 1. Having obtained the discretized form of the parameter ξ , we can obtain the correspond-
ing matrix Aξ j . To discretize the angular variable θ, we subdivide the interval [0, 2π ) into σ -pieces
of equal length and assume that θ0 = 2π /σ , σ ∈ N. Then, for this choice, we can write θ = θ0 ,
 ∈ Zσ = {0, 1, 2, . . . , σ − 1}. Finally, to facilitate a proper discretization of the translation param-
eter b, we shall take into consideration both the preceding discretizations of ξ and θ such that bn =
bjn0 n1 = B−j Rθ0 (n0 α0 , n1 α1 ) = B−j Rθ0 un0 n1 , where un0 n1 = (n0 α0 , n1 α1 ), n0 , n1 ∈ Z and α0 , α1 ≥
0. Therefore, the discretization grid is given by
  
    
G λ, σ , α0 , α1 = Bj 1, 1 , 2π /σ , bj n0 n1 , j, , n0 , n1 ∈ Z × Zσ × Z × Z . (33)


Corresponding to the grid (33), we define a discrete collection ψj,n (t) generated from a single
0 ,n1
window function ψ(t) ∈ L (R ) as
2 2

    


ψj,n (t) := det A j exp i t1 + t2 ψ R −θ A j t − un0 n1 , (34)
0 ,n1 ξ ξ
λj λj 0

where (j, , n0 , n1 ) ∈ Z × Zσ × Z × Z, and α0 , α1 ≥ 0. Formally, we have the following definition

of the discrete non-isotropic angular Stockwell transform.

Definition 4.1: For any function f ∈ L2 (R2 ), the discrete non-isotropic angular Stockwell transform
with respect to a window function ψ ∈ L2 (R2 ) is defined by

  
j
 1 
Mψ f bn , ξ , θ0 = f (t) ψj,n0 ,n1
(t) dt, (35)
2π R2


where ψj,n (t) is given by (34).
0 ,n1

For the illustration of the discrete non-isotropic angular Stockwell transform ( 35), we have the
following example.

Example 4.2: Consider the modulation function f (t) = eit·ω and the non-isotropic Gaussian win-
dow function ψ(t) = eik·t e−(1/2)|t| , k = (k0 , k1 ) ∈ R2 , t = (t1 , t2 ) ∈ R2 . To facilitate the computa-
2

tion of the proposed transform (35), we shall fix θ0 = 2π /4 so that the discretized parameter set
corresponding to θ becomes {0, π/2, π,  3π /2}. For our convenience, we fix the translation parameter
bn = (0, 0) and λ = 2, so that Aξ j = 2−j 0 . Consequently, the angular matrices take the form
0 2−j

       
1 0 0 −1 −1 0 0 1
R0 = , R−π/2 = , R−π = , R−3π/2 = .
0 1 1 0 0 −1 −1 0

Therefore, we have
    
  t1 t2 t2 t1
ψ R 0 Aξ j t − 0 = ψ , , ψ R −π/2 A ξ j t − 0 = ψ − , ,
2j 2j 2j 2j
     
  t1 t 2 t2 t1
ψ R−π Aξ j t − 0 = ψ − j , − j , ψ R−3π/2 Aξ j t − 0 = ψ ,− .
2 2 2j 2j
854 F. A. SHAH AND A. Y. TANTARY

Hence

  1  1
 −∞  −∞  
t1 t2

Mψ f 0, j , 0 = exp −i +
2 2π · 4j −∞ −∞ 2j 2j
× exp{i(t1 ω1 + t2 ω2 )}ψ(R0 Aξ j (t1 , t2 ) − 0) dt1 dt2
 −∞  −∞   
1 t1 t2
= exp −i j + j
2π · 4j −∞ −∞ 2 2
 
t1 t2
× exp{i(t1 ω1 + t2 ω2 )}ψ , dt1 dt2
2j 2j
 −∞  −∞   
1 t1 t2
= exp −i + exp{i(t1 ω1 + t2 ω2 )}
2π · 4j −∞ −∞ 2j 2j
    
k0 t1 k1 t2 (t12 + t22 )
× exp −i + exp − dt1 dt2
2j 2j 2 · 4j
 −∞    
1 1 k0 t12
= exp i − + ω 1 − t 1 − dt1
2π · 4j −∞ 2j 2j 2 · 4j
 −∞    
1 k1 t22
× exp i − j + ω2 − j t2 − dt2
−∞ 2 2 2 · 4j
 '    (
4j 1 k1 2 1 k0 2
= exp − ω2 − j − j + ω1 − j − j .
2 2 2 2 2

Similarly, we can have

 ' 2 (
  1 π  4j 1

k1 2

1 k0
Mψ f 0, j , = exp − ω1 − j − j + ω2 − j
+ j ,
2 2 2 2 2 2 2
 ' 2 (
  1  4j 1

k0 2

1 k1
Mψ f 0, j , π = exp − ω1 − j + j + ω2 − j
+ j ,
2 2 2 2 2 2

and
 '  (
   1 3π  4j 1

k1 2

1 k0 2
Mψ f 0, j , = exp − ω1 − j + j + ω2 − j − j .
2 2 2 2 2 2 2

We now compute the Fourier transform of the discretized family ψj,n  (t) given by (34), which will
0 ,n1
be invoked for establishing a sufficient condition for the underlying discrete family to be a frame for
L2 (R2 ).


Lemma 4.3: For the discrete family ψj,n (t) given by (34), the following identity holds:
0 ,n1

   

F ψj,n0 ,n1
(t) (ω) = exp{i(ξ j − ω) · A−1 R u
ξ j  θ0 n0 n1
}F ψ̃ (ωA−1 R
ξ j − θ0
), (36)

where ψ̃(t) = exp{i(1/λj , 1/λj ) · (A−1 R t)}ψ(t).

ξ j θ0
 APPLICABLE ANALYSIS 855

Proof: For t = (t1 , t2 ), ω = (ω1 , ω2 ) ∈ R2 , we have

  
 1
F ψj,n0 ,n1
(t) (ω) = ψ (t) e−iω·t dt
2π R2 j,n0 ,n1
      
| det Aξ j | 1 1 −iω·t
= exp i , · t ψ R −θ A
0 ξ j t − un0 1 e
n dt
2π R2 λj λj
    
1 1 1 −1
= exp i , · A R
ξ j θ0
(z + un0 n1 ) ψ(z)
2π R2 λj λj
× exp{−iω · A−1 R (z + un0 n1 )} dz
ξ j  θ0
    
1 1 1
= exp{i(ξ j − ω) · A−1 R u
ξ j θ0 n0 n1
} exp i , · (A −1
R
ξ j θ0
z) ψ(z)
2π R2 λj λj
× exp{−iω · A−1 R z} dz
ξ j θ0

= exp{i(ξ j − ω) · A−1 R u
ξ j θ0 n0 n1
}F
     
1 1 −1
× exp i , · (A R
ξ j  θ0
z) ψ(z) (ωA−1 R
ξ j − θ0
)
λj λj
= exp{i(ξ j − ω) · A−1 R u
ξ j  θ0 n0 n1
}F [ψ̃](ωA−1 R
ξ j − θ0
).

The proof of Lemma 4.3 is complete. 

Observation. From Lemma 4.3, we can deduce that


|F [ψj,n0 ,n1
(t)](ω)| = |F [ψ̃](ωA−1 R
ξ j −θ0
)|. (37)

The rest of the section is devoted to find conditions on the parameters λ, σ , α0 , α1 , so that the dis-

cretized family ψj,n (t) defined by (34) constitutes a frame for L2 (R2 ). In this direction, we have
0 ,n1
the following theorem.

Theorem 4.4: Suppose that a window function ψ ∈ L2 (R2 ) satisfies the following conditions:
σ
))
(i) P(λ, σ , ψ) = ess inf |F [ψ̃](ωA−1 R
ξ j −θ0
)|2 > 0,
ω∈R2
j∈Z =0
σ
))
(ii) Q(λ, σ , ψ) = sup |F [ψ̃](ωA−1 R
ξ j −θ0
)|2 < ∞,
ω∈R2 j∈Z =0
1+ε
(iii) sup (1 + |u|) β(u) < ∞, ε > 0
u∈R2

where
σ    
))
β(u) = sup −1 −1
F ψ̃ (ωAξ j R− θ0 + u) F ψ̃ (ωAξ j R− θ0 ) . (38)
ω∈R2 j∈Z =0

Then, there exist critical constants γ0 and γ1 such that for all α0 ∈ (0, γ0 ) and α1 ∈ (0, γ1 ), the
family (34) forms a frame for L2 (R2 ).
856 F. A. SHAH AND A. Y. TANTARY

* * 
Proof: Assume that I = j, n0 ,n1 ∈Z |f , ψj,n0 ,n1 | . Then, we observe that
2

) )   
 
I= f , ψj,n ,n
0 1
f , ψ j,n0 ,n1
j, n0 ,n1 ∈Z
) )    
 
= F f (ω), F [ψj,n0 ,n1
](ω)F f (η), F [ψj,n0 ,n1
](η)
j, n0 ,n1 ∈Z
) )  
     
= F f (ω)F f (η)F [ψj,n0 ,n1
](η)F [ψj,n0 ,n1
](ω) dω dη
j, n0 ,n1 ∈Z R2 R2

) )  
     
= F f (ω)F f (η){exp{−i(ξ j − ω) · A−1
ξ j R  θ 0 un0 n1 }F ψ̃ (ωA−1 R
ξ j − θ0
)}
j, n0 ,n1 ∈Z R R2
2

 
× {exp{i(ξ j − η) · A−1 R u
ξ j θ0 n0 n1
}F ψ̃ (ηA−1 R
ξ j − θ0
)} dω dη
) )        
   
= | det Aξ j |2 F f (ωR θ0 Aξ j )F f (ηR θ0 Aξ j )F ψ̃ ω F ψ̃ η
j, n0 ,n1 ∈Z R R2
2


× exp i(ω − η) · un0 n1 dω dη. (39)
Applying the Poisson’s summation formula
) )  
i(ω−η)·un0 n1 4π 2 ) )   2π 2π
e = δ ω − η − ũn0 n1 , with ũn0 n1 = n0 , n1
α0 α1 α0 α1
n0 ∈Z n1 ∈Z n0 ∈Z n1 ∈Z

so that

4π 2 | det Aξ j | ) )      
I= F f ωR θ0 Aξ j F f (ω − ũn0 n1 )R θ0 Aξ j
α0 α1
j, n0 ,n1 ∈Z R
2

     
× F ψ̃ ω F ψ̃ ω − ũn0 n1 dω

4π 2 ) )     
= F f ω F f (ω − ũn0 n1 R θ0 Aξ j )
α0 α1 R2 j, n0 ,n1 ∈Z
   
× F ψ̃ (ωA−1 ξ j R − θ 0 )F ψ̃ (ωA−1 R
ξ j − θ0
− ũn0 n1 ) dω
    2   2
4π 2 )
= F f ω F ψ̃ (ωA−1 ξ
R − θ ) dω
α0 α1
j 0
R 2
j,
) ) 
4π 2     
+ F f ω F f (ω − ũn0 n1 R θ0 Aξ j )
α0 α1
j, n0 ,n1 ∈Z\{0} R
2

   
× F ψ̃ (ωA−1
ξ j R − θ 0 )F ψ̃ (ωA−1 R
ξ j − θ0
− ũn0 n1 ) dω

= |I1 | + I2 . (40)
It is easy to deduce the estimate for I1 as
4π 2 P(λ, σ , ψ) # #
#  #2 4π 2 Q(λ, σ , ψ) # #
#  #2
#F f # ≤ |I1 | ≤ #F f # .
α0 α1 2 α0 α1 2
 APPLICABLE ANALYSIS 857

For obtaining the estimate for I2 , we use the Cauchy–Schwartz’s inequality so that

4π 2 ) )     
|I2 | ≤ F f ω F f (ω − ũn0 n1 R θ0 Aξ j )
α0 α1
j, n0 ,n1 ∈Z\{0} R
2

   

× F ψ̃ (ωA−1 ξ j R − θ 0 ) F ψ̃ (ωA −1
ξ j R − θ 0 − ũ n0 n1 ) dω
  
4π 2 ) )
≤ F ψ̃ (ωA−1 ξ
R − θ )
α0 α1
j 0
j, n0 ,n1 ∈Z\{0}R 2

     2 1/2

× F ψ̃ (ωA−1
ξj
R − θ 0 − ũn n )
0 1 F f ω dω
       2 1/2

× F ψ̃ (ωA−1 R
ξ j − θ0
) F ψ̃ (ωA−1
R
ξ j − θ0
− ũ n0 n1 ) F f (ω − ũn0 n1  θ0 ξ
R A j )
R2
4π 2 ) )
≤
α0 α1
j, n0 ,n1 ∈Z\{0}
        2 1/2

× F ψ̃ (ωA−1 R
ξ j − θ0
) F ψ̃ (ωA−1
R
ξ j − θ0
− ũ )
n0 n1 F f ω dω
R2
       2 1/2

× F ψ̃ (ωA−1 R
ξ j − θ0
+ ũn0 n1 ) F ψ̃ (ωA−1
R
ξ j − θ0
) F f (ω) dω .
R2

Again, by applying Cauchy–Schwartz’s inequality to the summation over j, , we get

⎧ ⎛
4π 2 ) ⎨ )  
⎝
|I2 | ≤ F ψ̃ (ωA−1 R
ξ j − θ0
)
α0 α1 ⎩ R2
n0 ,n1 ∈Z\{0} j,
⎞ ⎫1/2
     2 ⎬

× F ψ̃ (ωA−1 R
ξ j − θ0
− ũn0 n1 )⎠ F f ω dω
⎭
⎧ ⎛ ⎞ ⎫
⎨ )     2 ⎬1/2
  
× ⎝ F ψ̃ (ωA−1 R
ξ j − θ0
+ ũn0 n1 ) F ψ̃ (ωA−1 R
ξ j − θ0
)⎠ F f ω dω .
⎩ R2 ⎭
j,

Implementing the assumption (iii) of the hypothesis, we obtain

⎛ ⎞
4π 2 )     # #
#  #2
|I2 | ≤ ⎝ β ũn0 n1 β − ũn0 n1 ⎠ #F f #
α0 α1 2
n0 ,n1 ∈Z\{0}

4π 2 # #
#  #2 )   2
= #F f # β ũn0 n1
α0 α1 2
n0 ,n1 ∈Z\{0}

4π 2 #  #2  
# #
= #F f # H λ, σ , α0 , α1 , ψ , (41)
α0 α1 2

*
where H(λ, σ , α0 α1 , ψ) = n0 ,n1 ∈Z\{0} [β(ũn0 n1 )]
2 < ∞ in light of assumption (iii).
858 F. A. SHAH AND A. Y. TANTARY

Finally, to accomplish the proof, we shall use the following inequality:

|I1 | − |I2 | ≤ I = |I1 | + I2 ≤ |I1 | + |I2 |

to have
 
4π 2     # #
#  #2
P λ, σ , ψ − H λ, σ , α0 , α1 , ψ #F f # ≤ |I1 | − |I2 | (42)
α0 α1 2
  # #
4π 2     #  #2
and Q λ, σ , ψ + H λ, σ , α0 , α1 , ψ #F f # ≥ |I1 | + |I2 |. (43)
α0 α1 2

By assumptions (i)–(iii), it follows that P(λ, σ , ψ) is strictly positive and H(λ, σ , α0 , α1 , ψ) → 0 as

(α0 , α1 ) → (0, 0). Therefore, there exist critical constants γ0 , γ1 > 0 such that the first term in (42)
is dominant for all α0 < γ0 , α1 < γ1 and hence the L.H.S of (42) is strictly positive and yields a lower
frame bound. From inequalities (42) and (43), we immediately obtain the estimates for the frame
bounds. Hence, we conclude that the discrete family (34) constitutes a frame for L2 (R2 ).
This completes the proof of Theorem 4.4. 

Acknowledgments
The authors are deeply indebted to the anonymous referees’ for meticulously reading the manuscript, pointing out many
inaccuracies and giving several valuable suggestions to improve the initial version of the manuscript to the present stage.

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding
The first author is supported by SERB (DST), Government of India, under Grant No. EMR/2016/007951.

References
[1] Gröchenig K. Foundations of time-frequency analysis. Boston: Birkhäuser; 2001.
[2] Ali ST, Antoine JP, Gazeau JP. Coherent states, wavelets, and their generalizations. New York: Springer; 2015.
[3] Gabor D. Theory of communications. J Inst Electr Eng. 1946;93:429–457.
[4] Debnath L, Shah FA. Wavelet transforms and their applications. New York: Birkhäuser; 2015.
[5] Debnath L, Shah FA. Lecturer notes on wavelet transforms. Boston: Birkhäuser; 2017.
[6] Gomes J, Velho J. From Fourier analysis to wavelets. New York: Springer; 2015.
[7] Addison PS. The illustrated wavelet transform handbook. Boca Raton: CRC Press; 2017.
[8] Stockwell RG, Mansinha L, Lowe RP. Localization of the complex spectrum: the S transform. IEEE Trans Signal
Process. 1996;44:998–1001.
[9] Candés EJ, Donoho DL. Ridgelets: a key to higher-dimensional intermittency? Phil Trans R Soc London A.
1999;357:2495–2509.
[10] Candés EJ, Donoho DL. Continuous curvelet transform I: resolution of the wavefront set. Appl Comput Harmon
Anal. 2005;19:162–197.
[11] Do MN, Vetterli M. The contourlet transform: an efficient directional multiresolution image representation. IEEE
Trans Image Proc. 2005;14:2091–2106.
[12] Kutyniok G, Labate D. Resolution of the wavefront set using continuous shearlets. Trans Amer Math Soc.
2009;361:2719–2754.
[13] Prasad A, Manna S, Mahato A, et al. The generalized continuous wavelet transform associated with the fractional
Fourier transform. J Comput Appl Math. 2014;259:660–671.
[14] Stockwell RG. Why use the S transform? Pseudo-Differ Oper Partial Differ Equ Time Frequency Anal.
2007;52:279–309.
[15] Stockwell RG. A basis for efficient representation of the S-transform. Digit Signal Process. 2007;17:371–393.
[16] Du J, Wong MW, Zhu H. Continuous and discrete inversion formulas for the Stockwell transform. Integral
Transform Spec Funct. 2007;18:537–543.
 APPLICABLE ANALYSIS 859

[17] Wang Y, Orchard J. Fast discrete orthonormal Stockwell transform. SIAM J Sci Comput. 2009;31:4000–4012.
[18] Hutníková M, Mišková A. Continuous Stockwell transform; coherent states and localization operators. J Math
Phys. 2015;56:073504.
[19] Riba L, Wong MW. Continuous inversion formulas for multi-dimensional modified Stockwell transforms. Integral
Transform Spec Funct. 2015;26(1):9–19.
[20] Battisti U, Riba L. Window-dependent bases for efficient representations of the Stockwell transform. Appl Comput
Harmon Anal. 2016;40:292–320.
[21] Catanǎ V. Abelian and Tauberian results for the one-dimensional modified Stockwell transforms. Appl Anal.
2017;96(6):1047–1057.
[22] Saneva K, Atanasovaa S, Buralievab JV. Tauberian theorems for the Stockwell transform of Lizorkin distributions.
Appl Anal. 2018. doi:10.1080/00036811.2018.1506104
[23] Drabycz S, Stockwell RG, Mitchell JR. Image texture characterization using the discrete orthonormal S-transform.
J Digit Imaging. 2009;22(6):696–708.
[24] Moukadem A, Bouguila Z, Abdeslam DO, et al. A new optimized Stockwell transform applied on synthetic and
real non-stationary signals. Digit Signal Process. 2015;46:226–238.
[25] Liu Y, Wong M. Inversion formulas for two dimensional Stockwell transforms. Pseudo differential operators:
partial differential equations and time-frequency analysis; 2007, p. 323–330.
[26] Beckner W. Pitt’s inequality and the uncertainty principle. Proc Amer Math Soc. 1995;123:1897–1905.
[27] Folland GB, Sitaram A. The uncertainty principle: a mathematical survey. J Fourier Anal Appl. 1997;3:207–238.
[28] Dahlke S, Maass P. The affine uncertainty principle in one and two dimensions. Comput Math Appl.
1995;30:293–305.
[29] Battle G. Heisenberg inequalities for wavelet states. Appl Comput Harmonic Anal. 1997;4:119–146.
[30] Wilczok E. New uncertainty principles for the continuous Gabor transform and the continuous wavelet transform.
Doc Math. 2000;5:201–226.
[31] Guanlei X, Xiaotong W, Xiaogang X. The logarithmic, Heisenberg’s and short-time uncertainty principles
associated with fractional Fourier transform. Sig Process. 2009;89:339–343.
[32] Omri S. Logarithmic uncertainty principle for the Hankel transform. Integ Transf Spec Funct. 2011;22(9):655–670.
[33] Soltani F. Pitt’s inequality and logarithmic uncertainty principle for Dunkl transform on R. Acta Math Hungar.
2014;143(2):480–490.
[34] Mejjaoli H. Uncertainty principles for the generalized Fourier transform associated to a Dunkl-type operator.
Appl Anal. 2016;95(9):1930–1956.
[35] Hamadi NB, Omri S. Uncertainty principles for the continuous wavelet transform in the Hankel setting. Appl
Anal. 2018;97(4):513–527.
[36] Gumber A, Shukla NK. Uncertainty principle corresponding to an orthonormal wavelet system. Appl Anal.
2018;97(3):486–498.
[37] Su Y. Heisenberg type uncertainty principle for continuous Shearlet transform. J Nonlinear Sci Appl.
2016;9:778–786.
[38] Bahri M, Ashino R. Logarithmic uncertainty principle, convolution theorem related to continuous fractional
wavelet transform and its properties on a generalized Sobolev space. Int J Wavelets, Multiresol Informat Process.
2017;15(5):1750050 (22 pages).
[39] Shah FA, Tantary AY. Polar wavelet transform and the associated uncertainty principles. Int J Theor Phys.
2018;57(6):1774–1786.
[40] Cowling MG, Price JF. Bandwidth verses time concentration: the Heisenberg–Pauli–Weyl inequality. SIAM J Math
Anal. 1994;15(1):151–165.